Apr 16, 2010 - XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010 ... a classifier - Support Vector Machines algorithm: ⢠convergence to a ...
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Supervised Change Detection on Simulated Data employing Support Vector Machines
Ch. Psaltis, Ch. Ioannidis
XXIV FIG International Congress, Facing the Challenges – Building the Capacity Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Introduction
Automatic change detection of manmade objects Necessary for map or GIS update Gain time, effort, money
• Lots of research in the field • Success in certain very well-defined cases, by imposing certain constraints • Far from a turn-key solution
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
1
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Traits of change detection methods
(1/2)
Scale of the changes to be detected Urban area expansion, single building scale
Type of the basic comparison unit (compared feature) Grey tones, features, objects
Number of steps in which the process is completed Object extraction and comparison vs direct comparison
Path to change detection and a priori knowledge Bottom-up, top down approaches XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Traits of change detection methods
(2/2)
Deterministic (decisive answer as to what change is) or stochastic approach (a measure of how possible change is)
Level of automation Fully autonomous Automatic Semi automatic (training phase + prediction of changes)
Type of data used Raster, vector from different sources and sensors XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
2
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Proposed methodology for Change detection of manmade objects Supervised classification paradigm
Data for the same region in different time periods (orthoimages, DSMs) Evaluation of a number of predefined cues for the region Use of some manually collected positive and negative samples to train a classifier - Support Vector Machines algorithm: convergence to a global maximum requirement of a small number of training samples availability of good open source implementations
The classifier asserts change in the remaining data Tests of the proposed strategy with simulated data (validity and performance aspects of the strategy) XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Proposed method for buildings change detection
Single building scale
Orthoimages and DSMs from two measurement epochs
Direct comparison (one stage approach)
Bottom up technique, from pixels to changes
Semi-automatic, supervised classification
Deterministic method
Classifier: Radial Basis Function Support Vector Machines XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
3
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Workflow of the method 1. Acquisition of images from different periods + DSMs 2. Orientation of the images 3. Production of orthoimages for each period 4. Layering of the orthoimages & DSMs and calculation of the appropriate feature vectors 5. Training phase: manual labeling of positive and negative samples of changes 6. Classification of data as changed or not, using the trained SVM 7. Manual assessment of the resulted changes XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Support Vector Machines – Linear classifier SVM applications: from text recognition to image analysis Goal of SVM: to classify a given object to one of two available classes based on its characteristics
Binary classification Each object is viewed as N-dimensional vector Calculation of appropriate (N-1 d) hyperplane which best separates the two object classes: the distance to the nearest feature vector on each side is maximized Best separation of the training samples Use the classifier for new samples: classified depending on which side of the hyperplane they lay XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
4
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Linear SVM
Illustration of 2D Linear SVM from Burges (1998) XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Kernel functions
Non-linear hyperplanes
Better separation
Better accuracy
Common kernels Polynomial Radial basis function (RBF)
Illustration of non-linear hyperplane (DTREG, 2009)
SVM are a subclass of the kernel methods XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
5
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Testing on simulated data
Check the validity, effectiveness and robustness of the strategy
Information to improve the Selection of features, Feature space dimensionality, Fine tune feature selection Testing configuration of SVM parameters
Automated generation of data and classification Computer program in Python programming language
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Input for data generation
Buildings' average area, height and shape (buildings’ roof is considered flat)
Area of testing size and density in buildings
Distribution of buildings in the scene (grid or random type)
DTM with a uniform slope
Percentage of changes
Radiometric (N(m,σ)) and geometric noise (uniform translation, rotation, scaling) to the images
Height noise (N(m’,σ’)) of DSMs XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
6
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Output of data generation
Data for two measurement epochs
Greyscale simulations of orthoimages
Simulations of DSMs
Difference products from the above images and DSMs
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Data characteristics of the simulation tests
Area of interest: 1100 x 1000 pixel images
100 buildings: 75 old buildings (1st period) + 25 new buildings
Size of building: 200 pixels mean area
Shape of building: Orthogonal with arbitrary rotation and aspect ration of sides
Building height: 4m in average
Slope of original DTM: 10%
Size of cells to extract features: 10 x 10 pixels
Classification method: SVM
Testing scenarios:
Different buildings distributions (grid and random)
Data: Images or Images & DSMs
Different types of noise with varying magnitude XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
7
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Example of grid building distribution 1st period
2nd period + Geometric noise
Image Difference
Difference + Radiometric noise
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Example of random building distribution 1st period
2nd period + Geometric noise
Difference
Difference + Radiometric noise
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
8
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Testing procedure
In each scenario one train set and one test set are created
Generate differences and apply noise
Split area of interest in grid cells half the size of buildings
Calculate feature vectors (mean & standard deviation)
Automatically label changes for both datasets
The SVM is trained on the train set and then classifies the test set
The SVM also classifies the train set, to assess the validity of training
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Assessing results
Measurement of accuracy at object level: How many buildings were correctly detected as changed
Measurement of accuracy at grid cell level: True positive, true negative, false positive and false negative results
Sample difference image and grid cells XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
9
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
1. Classifying images with radiometric and geometric noise In each scenario the geometric noise is fixed to: 1 pixel displacement in x axis & 2 pixels in y axis 18 degrees rotation +10% scale in x axis & + 20% in y axis Radiometric noise varies (greyscale 0-255): mean noise 100 – 200 std noise 30 – 50 Results Objects accuracy: >96% True positives: >90% False positives: 0%
True negatives: 100% False negatives: 80% True positives: >67%
False positives: 90%) The final accuracy is proportional to the DSM quality XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
10
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
3. Classifying images and DSMs with different quality training and testing data Low noise training data High noise testing data Results DSM improves the results The final accuracy is proportional to the DSM quality: high quality DSM can greatly enhance the overall performance of the algorithm To achieve uniform levels of acceptable accuracy it is very crucial to train and test the SVM in almost the same conditions
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Conclusions
(1/2)
To improve the procedures for automated and reliable detection of new buildings, for several applications such as map updating, monitoring of informal development etc, a procedure using Support Vector Machines was developed and tested. The results using the Radical Basis Function SVM classifier are especially encouraging.
XXIV FIG International Congress, Sydney, Australia, 11-16 April 2010
FIG Congress 2010 Facing the Challenges – Building the Capacity Sydney, Australia, 11‐16 April 2010
11
05-05-2010
Laboratory of Photogrammetry School of Rural and Survey Engineering National Technical University of Athens
Conclusions
(2/2)
•
Applying simple noise conditions, even with high radiometric noise, the results are excellent All new buildings were detected and few false replies (positive or negative) were mentioned The proposed procedure works perfectly under ideal conditions
•
When the noise becomes complicated the success indicators are reduced (detection of new buildings 80% and 5% false negative responses) The use of DSMs seems to be necessary The proposed procedure gives the best results with DSMs of high quality and accuracy
•
When training and testing data differ in quality, the difficulties in detection of changes grow The use of images alone is no longer adequate (
